Example script stopped working

[andrew_peterson]

The example script with force training (e.g., here), is not converging as of the latest development version.

I tried it with 0.6.0 and it works fine, converging in about 300 steps. Things to check:

• Scipy version (Amp 0.6.0 test used 0.19.1 with python2, which is the slow one; newer Amp tests used Scipy v1.1.0 with python3)
• Default Gaussian fingerprint parameters have changed.
• Anything else?

[Muammar El Khatib]

It worked for me, using the script you mentioned. Attached you have the -log.txt using the plot-convergence tool.

"""Simple test of the Amp calculator, using Gaussian descriptors and neural
network model. Randomly generates data with the EMT potential in MD
simulations."""

import os
from ase import Atoms, Atom, units
import ase.io
from ase.calculators.emt import EMT
from ase.lattice.surface import fcc110
from ase.md.velocitydistribution import MaxwellBoltzmannDistribution
from ase.md import VelocityVerlet
from ase.constraints import FixAtoms

from amp import Amp
from amp.descriptor.gaussian import Gaussian
from amp.model.neuralnetwork import NeuralNetwork
from amp.model import LossFunction


def generate_data(count, filename='training.traj'):
    """Generates test or training data with a simple MD simulation."""
    if os.path.exists(filename):
        return
    traj = ase.io.Trajectory(filename, 'w')
    atoms = fcc110('Pt', (2, 2, 2), vacuum=7.)
    atoms.extend(Atoms([Atom('Cu', atoms[7].position + (0., 0., 2.5)),
                        Atom('Cu', atoms[7].position + (0., 0., 5.))]))
    atoms.set_constraint(FixAtoms(indices=[0, 2]))
    atoms.set_calculator(EMT())
    atoms.get_potential_energy()
    traj.write(atoms)
    MaxwellBoltzmannDistribution(atoms, 300. * units.kB)
    dyn = VelocityVerlet(atoms, dt=1. * units.fs)
    for step in range(count - 1):
        dyn.run(50)
        traj.write(atoms)


generate_data(20)

calc = Amp(descriptor=Gaussian(),
           model=NeuralNetwork(hiddenlayers=(10, 10, 10)))
calc.model.lossfunction = LossFunction(convergence={'energy_rmse': 0.02,
                                                    'force_rmse': 0.02})
calc.train(images='training.traj')

 5762 2018-08-15T14:09:46   3.2073e-04 5.4964e-05 C 1.3865e-04 C 2.0021e-02 - 1.3935e-01 C
 5763 2018-08-15T14:09:46   3.2056e-04 5.1715e-05 C 1.3805e-04 C 2.0016e-02 - 1.3936e-01 C
 5764 2018-08-15T14:09:46   3.2039e-04 5.2613e-05 C 1.3609e-04 C 2.0011e-02 - 1.3896e-01 C
 5765 2018-08-15T14:09:46   3.2025e-04 5.3339e-05 C 1.2897e-04 C 2.0006e-02 - 1.3876e-01 C
 5766 2018-08-15T14:09:46   3.2011e-04 5.1133e-05 C 1.2265e-04 C 2.0002e-02 - 1.3898e-01 C
 5767 2018-08-15T14:09:47   3.1997e-04 5.2586e-05 C 1.3098e-04 C 1.9997e-02 C 1.3936e-01 C
...optimization successful. 17.9 min.
Amp successfully trained. Saving current parameters.
Parameters saved in file "amp.amp".
This file can be opened with `calc = Amp.load('amp.amp')`

This is my OS info:

User: muammar
Hostname: nuc
Date: 2018-08-15T13:51:38 (2018-08-15T17:51:38 UTC)
Architecture: x86_64
PID: 29787
Amp version: 0.7.0-beta
Amp directory: /home/muammar/brown/git/amp/amp
 Last commit: 82d1a78ff265a8fc7446a6f259a2a28e0b97bca9
 Last commit date: 2018-08-14 14:50:35 +0000
Python: v3.6.6: /usr/bin/python3
ASE v3.16.3b1: /home/muammar/brown/git/ase/ase
NumPy v1.15.0: /home/muammar/.local/lib/python3.6/site-packages/numpy
SciPy v1.1.0: /home/muammar/.local/lib/python3.6/site-packages/scipy
ZMQ/PyZMQ v4.1.6/v17.0.0: /home/muammar/.local/lib/python3.6/site-packages/zmq
pxssh: Not available from pxssh.
pxssh (via pexpect v4.6.0): /home/muammar/.local/lib/python3.6/site-packages/pexpect/pxssh.py